Binning? Or cutting parts that they shouldn't like they did with the 970? One never knows.
You seem to be moving goal posts. The 970 was a bad design choice with its memory subsystem. Nothing like that has ever happened since.
This doesn't have much to do with binning, I guess.
So where are the fully enabled GB202 dies?
I think they'll never exist, just like the AD102 full die never came into light, simply because they're too big and yields are not known to reach 100% for such big dies.
I had already spoken about this one before.
And the rest that's coming out later in the 5060 and 5050?
5060ti will likely be the full GB206, whereas the 5060 will be a die cut out of that. The 5050 is rumored to be the full GB207.
What's the point you're trying to make with such examples?
If a refresh comes later, chances are, you'll be able to achieve the same clock speed with a little tuning.
It really depends, as we have seen before with the 480/580 vs the 590, it was the same die on a better node, so better clocks and efficiency overall.
They would also change the memory subsystem, which is not something a 9070xt owner will be able to achieve.
Also, Microcenter had 3100 vanilla 9070s at launch and 8500 XTs. So apparently, yields are good. If yields are good, then where are the fully enabled Nvidia GPUs?
Well, you said yourself, the 5080 uses the full GB203 die, and that's a bit larger than the Navi 48.
The GB202 is 2.1x bigger than Navi 48.
Those will be the Super cards, like they have been with Ada and Ampere as well. They just have to show how much more value you're getting, boom, profit.
The 5090 won't have any refresh, more than likely.
If a refresh of the 5080 happens, they'll only change the memory subsystem, or move it to be a die cut of the GB202 die.
5070 is likely to have a refresh, yeah.
Don't think the same will happen to the smaller models tho.
like they have been with Ada and Ampere as well
Only product that didn't use the full die at launch and got a refresh later was the AD103. And even then, the 4080 Super did not bring much to the table over the 4080 perf-wise, since it only enabled 4 SMs. The price drop was the thing that made it an interesting product.
Fun fact: the 4080 Super was the only product using the full AD103, not even enterprise products got such configuration, which does lead me to believe that the harvests for such die were not good enough for such product at launch.
I believe the above specially applies to Ampere's 3070ti and 3090ti, given how Samsung's node was a mess (albeit cheap).
It's not like Nvidia is doing this for the first time and I'm making it up. No, they've been doing this for the last 3 generations. Yet, people like you still don't believe it.
I just don't see your point. As I said, it's not like they're withholding good, fully enabled dies. They either sell those, or cut those down to segment a product, as any other company.
Refreshes may happen, companies do it all the time, specially as node maturity increases. Are you saying that new products should
not be launched?
Exactly. I'd be happy to see some unification attempts on that front, kind of like how all desktop CPUs are x86.
Not gonna happen, each GPU has its own specific ISA and higher level APIs.
Your best bet would be one of the higher level APIs supporting it, such as Vulkan. Nvidia has already contributed an extension that abstracts those that has been adopted by the Khronos group:
Machine learning harnesses computing power to solve a variety of ‘hard’ problems that seemed impossible to program using traditional languages and techniques. Machine learning avoids the need for a…
developer.nvidia.com
Seems like Valve has added support for it recently in Mesa for RDNA4:
www.phoronix.com
I'll also try FSR 4 once I get my card running properly in Linux. 
